cs.CVMar 19, 2026
SaveWhich Workloads Belong in Orbit? A Workload-First Framework for Orbital Data Centers Using Semantic Abstraction
Organizations: Independent Researcher, Twin Falls, ID, USA
Abstract
Space-based compute is becoming plausible as launch costs fall and data-intensive AI workloads grow. This paper proposes a workload-centric framework for deciding which tasks belong in orbit versus terrestrial cloud, along with a phased adoption model tied to orbital data center maturity. We ground the framework with in-orbit semantic-reduction prototypes. An Earth-observation pipeline on Sentinel-2 imagery from Seattle and Bengaluru (formerly Bangalore) achieves 99.7-99.99% payload reduction by converting raw imagery to compact semantic artifacts. A multi-pass stereo reconstruction prototype reduces ~306 MB to ~1.57 MB of derived 3D representations (99.49% reduction). These results support a workload-first view in which semantic abstraction, not raw compute scale, drives early workload suitability.
Figures & tables
Fig. 1: Space–ground connectivity scenarios and latency drivers.
Fig. 2: Bandwidth symmetry (illustrative).
| Dimension | Meaning (what it captures) |
|---|---|
| Latency tolerance | Sensitivity to added RTT, contact windows, and store-and-forward delays. |
| Bandwidth intensity | Expected reduction from raw inputs to downlinkable outputs; higher reduction improves suitability under downlink constraints. |
| Fault tolerance | Acceptability of retries, partial degradation, and intermittent execution without safety or mission impact. |
| Data locality | Whether the workload’s primary data is generated in space and loses value if moved in raw form. |
| Compute intensity | Computation required per unit of data transferred; higher compute per byte better amortizes constrained links. |
TABLE I: Evaluation criteria for workload suitability
| Score | Interpretation | Example |
|---|---|---|
| 1 | Real-time response ( s) | Collision avoidance, attitude control |
| 2 | Seconds | Dynamic retasking during pass |
| 3 | Tens of seconds | Near-real-time alerts |
| 4 | Minutes | Batch EO inference |
| 5 | Hours+ | Archival analytics |
TABLE II: Latency tolerance scoring (illustrative)
| Score | Interpretation | Example |
|---|---|---|
| 1 | Low reduction ( ) | Raw telemetry relay |
| 2 | Minor reduction (2–5 ) | Light compression |
| 3 | Moderate reduction (5–20 ) | Image downsampling |
| 4 | High reduction (20–100 ) | Segmentation masks |
| 5 | Extreme reduction ( ) | Polygons, object lists |
TABLE III: Bandwidth intensity scoring (illustrative)
| Score | Interpretation | Example |
|---|---|---|
| 1 | Mission-critical | Onboard control loops |
| 2 | Low tolerance | Navigation updates |
| 3 | Moderate tolerance | Event detection |
| 4 | High tolerance | Image filtering |
| 5 | Very high tolerance | Batch ML inference |
TABLE IV: Fault tolerance scoring (illustrative)
| Score | Interpretation | Example |
|---|---|---|
| 1 | Earth-originated data | Software updates |
| 2 | Mostly terrestrial | Ground sensor fusion |
| 3 | Mixed | Cross-domain analytics |
| 4 | Primarily space-generated | EO imagery |
| 5 | Exclusively space-native | RF spectrum capture |
TABLE V: Data locality scoring (illustrative)
| Score | Interpretation | Example |
|---|---|---|
| 1 | Very low | Packet forwarding |
| 2 | Low | Format conversion |
| 3 | Moderate | Image enhancement |
| 4 | High | CNN inference |
| 5 | Very high | Multi-stage ML pipelines |
TABLE VI: Compute intensity scoring (illustrative)
| Workload category | Lat. | BW | Fault | Local. | Comp. | Avg. |
|---|---|---|---|---|---|---|
| 3D reconstruction from satellite imagery | 4 | 4 | 4 | 4 | 5 | 4.2 |
| Space RF signal processing | 3 | 5 | 4 | 5 | 5 | 4.4 |
| EO preprocessing (radiometric, geometric) | 5 | 4 | 4 | 5 | 3 | 4.2 |
| Satellite health monitoring / telemetry analytics | 2 | 2 | 2 | 4 | 3 | 2.6 |
| Batch LLM inference | 4 | 2 | 4 | 2 | 4 | 3.2 |
| LLM training | 5 | 1 | 3 | 1 | 5 | 3.0 |
TABLE VII: Workload suitability matrix (scores 1–5 across criteria, illustrative)
| Workload | P1 | P2 | P3 | Rationale |
| 3D reconstruction from satellite imagery | ✓ | ✓✓ | P1: partial steps (tiling, feature extraction, limited depth); P2: cheap power enables full MVS/depth/mesh pipelines; P3: LISL enables in-space aggregation of multi-angle imagery. | |
| Space RF signal processing & classification | ✓ | ✓✓ | ✓✓ | P1: dense space-native RF benefits from detection/classification; P2: cheap power enables richer models; P3: LISL enables global spectrum maps and cross-orbit emitter tracking. |
| EO preprocessing (radiometric, geometric) | ✓ | ✓✓ | ✓✓ | P1: strong bandwidth reduction from space-native imagery; P2: higher duty cycle enables deeper pipelines and broader coverage; P3: LISL enables cross-satellite fusion and consistent processing. |
| Satellite health monitoring / telemetry analytics | P1: useful for autonomy, not bandwidth savings; P2: cheap power does not change limited data-reduction economics; P3: LISL enables fleet-level aggregation but remains secondary. | |||
| Batch LLM inference | × | ✓ | P1: weak ROI without cheap power and curated inputs; P2: economical for summarization/alerting on down-selected data; P3: aggregates outputs from multiple sources for semantic fusion. | |
| LLM training | × | × | P1: power and data gravity make it infeasible; P2: cheap power alone does not solve dataset ingress, hardware refresh, and tooling; P3: may support specialized space-derived training. |
TABLE VIII: Workload suitability and phase-wise roadmap (illustrative)
Fig. 3: Cross-region semantic abstraction across cloud regimes (Seattle vs Bengaluru, 2025).
Fig. 4: Example EO semantic artifacts (Seattle 2025).
Fig. 5: Multi-pass geometric abstraction pipeline (Los Angeles).